ZAP

ZAP (Zero-real Agent Policy) is a robot policy method that uses a pretrained vision-language model to provide world understanding and uses a simulation environment to teach action skills, without using any real robot data throughout.

ZAP achieves a 96.7% success rate on real-world tasks.

ZAP uses a small off-the-shelf 27B-parameter vision-language model and controls the robot through a tool-calling framework.

ZAP's training pipeline has two steps: first, bootstrapping world understanding and reasoning capability with a pretrained model, then teaching the model to move the robot on simple simulation tasks through supervised fine-tuning and GRPO.

Under the same tool framework, ZAP performs on par with GPT-6 Astra but runs 3 times faster and costs 17 times less.

Compared with the dedicated vision-language-action model MolmoAct2 VLA, ZAP's success rate is significantly higher (MolmoAct2 VLA's success rate is only 33.3%).

On the single-arm Franka robot, ZAP achieves 55.9% on the RoboLab-120 benchmark, leading the state-of-the-art VLA/WAM baseline by 4.9 percentage points.

ZAP can complete compositional generalization tasks that do not appear in the training data, including placing test tubes into a test tube rack and selecting hole positions, handing over test tubes between two arms in mid-air, drawing X on a tic-tac-toe board, and playing chess.

ZAP's core claim — that simulation training plus pretrained understanding is sufficient to support real robot control — currently comes from only a single source (rss), the 96.7% success rate was achieved on a specific task set and under a specific tool framework, and the original text does not elaborate on whether the task complexity and environmental diversity are sufficient to represent real deployment scenarios.

ZAP proposes a reproducible recipe: use pretrained models to solve the "what", use simulation to solve the "how", and replace end-to-end large models with small-scale off-the-shelf models plus a tool harness.